Ecography 36: 864–867, 2013
doi: 10.1111/j.1600-0587.2013.00321.x
Inference from presence-only data; the ongoing controversy
Trevor Hastie and Will Fithian­
T. Hastie (hastie@stanford.edu) and W. Fithian, Statistics Dept, Stanford Univ., CA 94305, USA.­
Presence-only data abounds in ecology, often accompanied by a background sample. Although many interesting aspects
of the species’ distribution can be learned from such data, one cannot learn the overall species occurrence probability, or
prevalence, without making unjustified simplifying assumptions. In this forum article we question the approach of Royle
et al. (2012) that claims to be able to do this.
Modelling of species distributions is most convincing
when presence–absence data is sampled in a systematic way.
For example, researchers survey a collection of equal sized
quadrats and record the presence or absence of a particular
species of plant. They also record other features of the
quadrat, such as annual precipitation, soil salinity, altitude,
and so on. These features are then used in a statistical model
such as logistic regression to build a model for the prob­
ability of species occurrence. Using this fitted model, species
occurrence probability is predicted and can be projected
onto a map of the region, if the features are available at
each geographical unit. See Guisan and Zimmerman (2000)
for a review of such methods.
Often the only species data available are the geographical
coordinates of sites where the species was observed –
so-called presence-only data – as recorded by observers.
Also available is a large collection of background data, con-
sisting of geographical coordinates and associated geographi-
cal features such as those available from GIS data. In many
cases this background data is available at every geographical
unit of area in a map of the region, and hence also at the
presence sites. Apart from those locations where the species
were observed, no species information is available for the
background data.
For animal species, sampling time as well as area is rele-
vant, since the species may wander around. Other compli-
cating factors exist, such as the species being present but are
not observed, and sampling bias (e.g. proximity to roads).
For the purposes of this article, we keep the discussion
simple and avoid these other sampling issues.
The question is what can one learn from such presence–
background data? There are many approaches to this prob-
lem, which has become an active area of research. Our
current favorite approach is to model the species occurrence
rate (as in number of times the focal species is seen per unit
area, per unit time). To this end the inhomogeneous Poisson
point process (IPP) (Wharton and Shepard 2010, Aarts et al.
864
© 2013 The Authors. Ecography © 2013 Nordic Society Oikos
Subject Editor: Miguel Araújo. Accepted 22 February 2013
2012) is attractive. Other popular approaches are: 1)
MAXENT (Phillips and Dudik 2008) which models the fea-
ture density for the presence data. 2) Naive logistic regres-
sion, which treats the background data as absence data, and
fits logistic regression models. 3) Manly’s exponential model
(Manly et  al. 2010), a precursor to the IPP model in this
context. Fithian and Hastie (unpubl.) survey these methods,
and shows that they are all equivalent to the IPP model, in
particular for an exhaustive sample of background data.
Similar conclusions appear in Aarts et al. (2012) as well as
Warton and Shepherd (2010). The conclusion to be drawn
from these comparisons is that while absolute occurrence
rates are typically elusive, relative rates are more accessible
and available from these type of data.
This forum article will step away from these sampling
considerations, and address a simpler question. We can think
of sampling units as geographical sites x ∈ c; each x repre-
sents a cell or unit of area, and c represents the domain of
interest, or entire collection of such cells. At each site the
vector z  z(x) records the values of som geographical attri-
butes or features. Let the binary variable y denote presence
(1) or absence (0) sites. We denote the marginal density of z
by p(z), and the (conditional) density of z at presence sites
by p1(z), and absence sites by p0(z). If the overall presence
occurrence probability is y(y  1) and hence absence
y(y  0)  1 2 y(y  1), then basic probability theory tells
us two things:
1) the conditional occurrence probability at a site, given
we observe feature z, is given by
ψ( y  1) π1( z )
ψ( y  1z )  (1)
π( z )
2) The marginal feature density p(z) is a mixture of the two
class-conditional densities:
p(z) 5 p1(z) y(y 5 1) 1 p0(z)(1 2 y(y 5 1)) (2)
 Presence–background data consists of a random sample
of values of z from p1(z), as well as a separate sample from
p(z) (possibly the entire background distribution), which
directly inform us about the densities p1 and p. However,
even if both of these distributions were fully known, we can
see from Eq. (1) that this would not be enough information
to estimate y(y  1z). We are missing the overall occurrence
probability y(y  1), or at least some data that allow us to
estimate this.
The reader might think that Eq. (2) offers some hope, but
it does not. We know or have data on p(z) and p1(z) – this
leaves a lot of flexibility in choosing somewhat arbitrary
values for y(y  1) and p0(z) to make Eq. (2) work out –
unless, that is, we impose strong parametric restrictions on
some of the ingredients. But then we are manufacturing
information via these assumptions when none exists in
the data. We will see more of this in the next section. Ward
et  al. (2009) discuss this problem and the lack of identifi-
ability of y(y  1) from such data. They warned of the folly
in relying on arbitrary parametric assumptions to squeeze
out estimates of y(y  1). Phillips et al. (2009) raise similar
issues. Most recently Phillips and Elith (2013) address the
same issue, and reinforce some of the points we make here.
The parametric approach of Royle et al.
Royle et  al. (2012) discuss methods for estimating species
occurrence probabilities from presence-only data – the same
problem we outline above. They cite Ward et al. (2009), yet
proceed to impose parametric assumptions on y(y  1z) to
enable estimation of y(y  1) – exactly what we warned
against. Here we will strengthen our argument in the context
of their model, and using their notation. We will also sim-
plify the discussion further, as they did, and focus attention
on geographic features x rather than environmental features
z  z(x); in the appendix we show that this transition is
benign.
Figure 1 (left panel, red) shows a plot of a very simple
model for occurrence probability
y(y  1x; b) (3)
This corresponds to a logistic regression model linear in x,
logit[y(y 1x; b)]  b0 1 xb1 (4)
In this case b0  21 and b1  1. We assume here that the
marginal distribution p(x) is uniform on [22.5, 2.5],
which makes the overall prevalence y(y  1)  ∫y(y   1x; b)
p(x)dx  ≈  0.33 in this case, and the values of y(y  1x; b)
range between 0.03 and 0.83. Royle et al. (2012) use a linear
logistic model similar to Eq. (4) for modeling occurrence
probability.
With such a parametric assumption, one can perform
inference on the data. As they point out, using Eq. (1)
we can write
ψ( y  1) π1( x )  ψ( y  1x ) π( x ) (5)
 y( y  1, x) (6)
If p(x) is uniform, as it typically is in the geographic domain,
and since Σx ∈c y(y  1, x)  y(y  1), we can write
ψ( yi  1xi )
π1( xi )  (7)
Σ x∈χ ψ( y  1x )
This is a model for the density of the observed data xi
at the presence sites, and it is expressed in terms of the
parameters of our logistic regression model if we replace
y(yx) in Eq. (7) with y(yx; b). On the basis of this
Royle et al. (2012) do maximum-likelihood estimation for
b see Eq. (9) below. Note that the presence observations xi
appear in the numerator; the background data are used to
compute the sums in the denominator.
This sounds like statistical alchemy: why don’t we need to
know y(y  1) anymore? Note that b0 is playing a similar

Occurrence probability Logit of occurrence probability

1.0 ψ(y = 1|x; β)
ψ∗(y = 1|x; β) 1
0.8
0
logit[Pr(y = 1|x)]
Pr(y = 1|x)

0.6 −1

0.4 −2

0.2 −3

0.0 −4

−2 −1 0 1 2 −2 −1 0 1 2
χ χ

Figure 1. Left panel: two different models for occurrence probabilities. The blue curve is half the red curve, and hence has exactly half
the prevalence (marginal occurrence probability) of the red curve. The implied likelihood (Eq. 9) of the presence-data xi is identical for these
two models. Right plot: the logits of the two models on the left. The broken blue curve is the best linear approximation to the solid blue
curve – the approximation that would be imposed by a linear logistic regression model. Since the solid logit curves (red and blue)
are indistinguishable with respect to the likelihood (Eq. 9), distinguishing the dotted blue line from the red is no easier than distinguishing
it from the solid blue. Determining whether or not this slight curvature is present is the entire basis upon which the Royle et al. (2012)
procedure would estimate the prevalence at either 34% (solid red) or 17% (dotted blue).

865
role as y(y  1) was before (y(y  1) multiplied all occur-
rence probabilities, whereas eb0 multiplies the odds).
Therefore, it should be surprising that suddenly we can esti-
mate it from the data. The reason b0  21 vs b0  1 are
distinguishable from each other in this model is that if
we hold b1 fixed and change b0, it increases all of the prob-
abilities y(y  1x) in Eq. (7), increasing the numerator and
the denominator. But because it changes some a little more
than others, it subtly changes the density p1(xi) – ‘subtly’
being the operative word. The problem is that other things
can subtly change p1 too – such as the linear logistic model
(Eq. 4) being subtly misspecified, as we see next.
The blue curve in this left hand plot of Fig. 1 corresponds
to a different model for the occurrence probability,
1 the likelihood Eq. (9), distinguishing the dotted blue
ψ ∗( y  1x ; β)   ψ( y  1x ; β)
2 line from the red is no easier than distinguishing it from
For this model (Eq. 8) the overall prevalence y*(y  1)  
1/2  y(y  1), i.e. 0.17 or exactly half of the prevalence
for model (Eq. 3). Although y*(y  1x) does not corre-
spond to a linear logistic model, it is nearly linear on
the logit scale (see the solid blue curve in the right plot of
Fig. 1), and is still a simple parametric model; but more
on that to come.
The critical point of this example is that the joint likeli-
hood of the presence data (i.e. Eq. 4 in Royle et al. (2012))
n ψ( yi  1x i ; β)
£( β) ∏
i 1 Σ x∈χ ψ( y  1x ; β)
n ψ ∗ ( yi  1x i ; β)
∏
i 1 Σ x∈χ ψ ∗ ( y  1x ; β)
is identical for these two models. In other words, the
likelihood would have nothing to say about whether model
(Eq. 3) or model (Eq. 8) was preferred – two models, both
with two parameters, but with one having prevalence half
the other. We could change the 1/2 in (Eq. 8) to any 0  
C  1 and the same statement would be true (with ‘half ’
changed to ‘fraction C ’). We note that this lack of identifi-
ability with propotional models that we exploit was pointed
out by Lele and Keim (2006, p. 3023, top left), who origi-
nally proposed the approach used by Royle et al. (2012).
Now the second model is not a linear logistic model, so it
would not be up for comparison in the Royle et al. (2012)
framework. The right hand panel shows the logit transforms
of each of these two models. Indeed, the second model is
not a linear logistic model, but it is almost one. The dotted
blue curve shows the best linear approximation to this logit
in the population. Since the solid red line and solid blue
curve are indistinguishable from each other with respect to
(8)
the solid blue. Determining whether or not this slight
curvature is present is the entire basis upon which the
Royle et  al. (2012) procedure would estimate the preva-
lence at either 34% (solid red) or 17%. One would need an
awfully large amount of data to be able to detect a difference
between the two blue lines, even with presence–absence
data.
We now present a simulation to reinforce the points
we have made. We simulate data from model (Eq. 8)
(nearly linear logistic), and fit a linear logistic model using
the likelihood (Eq. 9). In detail, we generate a large sample
of values of x via the uniform distribution p, generate
(9) 0/1 ‘presence/absence’ data using the probabilities (Eq. 8),
and then take a random subset of 1000 values of x from
those that came up as ‘present’. This sample of 1000 is
(10) fed into (Eq. 9), which is then maximized with respect to
the two linear logistic parameters. Rather than show the
parameter estimates b̂ 0 and b̂ 1 that result, we instead com-
pute the implied estimated prevalence value ŷ(y  1)  
∫y(y  1x; b̂)p(x)dx. This was repeated B  1000 times.
The middle histogram in Fig. 2 shows the results. The

C= 0.3 C= 0.5 C= 0.7

200 200 200

150 150 150

Frequency
Frequency

Frequency

100 100 100

50 50 50

0 0 0
0.0 0.1 0.2 0.3 0.4 0.0 0.1 0.2 0.3 0.4 0.0 0.1 0.2 0.3 0.4
ψ(y = 1) ψ(y = 1) ψ(y = 1)

Figure 2. Results of three separate simulation runs. The center histogram shows the estimated value ŷ(y  1) when a linear logistic
regression model is fit to data generated from model (Eq. 8). The histogram summarizes B  1000 different runs, each consisting of
1000 presence samples. The true value of y(y  1) is given by the vertical red line. The two flanking histograms change the 1/2 in
(Eq. 8) to C  3/10 (left) and C  7/10 (right), with again the red line showing the true value of y(y  1). In all cases, irrespective of
the true value of y(y  1), the histograms indicate that the value being estimated is centered around y(y  1) ≈ 0.33, the value from
Eq. (3).

866
histogram is peaked around 0.33 (the value from Eq. (3), not
the true value 0.17). The flanking histograms repeat these
simulations using 3/10 and 7/10 rather than the 1/2 in
(Eq. 8). In all cases the estimated ŷ(y  1)bear no relation-
ship to the true values (red lines).
The take-home message here is that: a) two perfectly
good and parsimonious probability models are indistin-
guishable with respect to the likelihood (Eq. 9) for the pres-
ence data, despite the fact that one has half the prevalence
of the other; i.e. prevalence is not identifiable in this
extended family. b) By insisting on a particular parametric
form, e.g. linear logistic, we are on extremely flimsy ground,
as the subtle distinction in this example shows. c) When
presence-only data arise via models that are nearly linear
on the logistic scale, maximum likelihood using Eq. (9)
and a linear logistic regression model can be incapable of
estimating the correct parameter values, and in particular
the correct implied prevalence. This is not trickery with data
simulations or abstract ideas; it cuts to the core of how
Royle et al.’s model estimates probabilities. They say you can
estimate probabilities from presence-only data by using
their model, which relies on a linear logistic framework.
The problem is that in the real world, functional forms
are almost never linear; linearity is just a useful approxi­
mation. We have shown here that data distributed just
slightly differently to that allowed in their framework will
lead to incorrect estimates of prevalence, and therefore
incorrect estimates of probability of presence.
Stepping back a bit, we remake our earlier point. It
should be clear that a sample of n1 sites, along with a
sample of n0 unclassified samples (i.e. a mix of presence and
absence), tells you nothing about the overall probability of
occurrence, absent strong parametric assumptions about
the form of the underlying densities. In other words, there is
no information on prevalence in the data itself; it all comes
from the model assumptions. Using such assumptions as the
basis for estimating overall prevalence is not a good idea; as
Appendix
Geographic vs environmental features
Usually we parameterize our conditional occurrence model
in terms of environmental features zi  z(xi) rather than
the geographic features xi themselves. For ease of exposition
here, we treat z as discrete rather than continuous.
Then along the lines of Eq. (7) we would have
ψ( yi  1zi ) π( zi )
π1( zi )  (11)
Σ z ∈Z ψ( y  1z ) π( z )
Here p(z) is the marginal environmental feature distribution,
and is not uniform. However,
π( z )  ∑ π( x ) (12)
x∈χ ; z ( x )  z
and so the denominator in Eq. (11) can be written
shown here, they are too fragile and arbitrary, and will not be
robust in practical settings.­­
Acknowledgements – TH was partially supported by grant
DMS-1007719 from the National Science Foundation, and
grant RO1-EB001988-15 from the National Inst. of Health. WF
was supported by VIGRE grant DMS-0502385 from the National
Science Foundation. The authors thank Jane Elith for helpful sug-
gestions on an earlier draft.
References
Aarts, G. et al. 2012. Comparative interpretation of count,
presence–absence and point methods for species distribution
models. – Methods Ecol. Evol. 3: 177–187.
Guisan, A. and Zimmerman, N. 2000. Predictive habitat dis­
tribution models in ecology. – Ecol. Model. 135: 147–186.
Lele, S. and Keim, J. 2006. Weighted distributions and estimation
of resource selection probability Functions. – Ecology 87:
3021–3028.
Manly, B. et  al. 2010. Resource selection by animals: statistical
design and analysis for field studies. – Kluwer.
Phillips, S. and Dudik, M. 2008. Modeling of species distribution
with maxent: new extensions and a comprehensive evaluation.
– Ecography 31: 161–175.
Phillips, S. and Elith, J. 2013. On estimating probability of
presence from use-availability or presence-background data.
– Ecology in press.
Phillips, S. et al. 2009. Sample selection bias and presence-only
distribution models: implications for background and pseudo-
absence data. – Ecol. Appl. 19: 181–197.
Royle, J. et  al. 2012. Likelihood analysis of species occurrence
probability from presence-only data for modelling species
distributions. – Methods Ecol. Evol. 3: 545–554.
Ward, G. et  al. 2009. Presence-only data and the em algorithm.
– Biometrics 65: 554–563.
Wharton, D. and Shepard, L. 2010. Poisson point process models
solve the ‘pseudo-absence problem’ for presence-only data
in ecology. – Ann. Appl. Stat. 4: 1383–1402.
∑ ψ ( y 1z) π( z ) ∑ ψ ( y 1z( x )) π( x ) (13)
z ∈Z x ∈χ
So then we have
ψ ( y i  1z ( xi ))
π1 ( zi )   π( zi ) (14)
∑ x ∈χ ψ ( y  1z ( x ))π( x )
So if p(x) is uniform, and we parameterize y(yi  1z(xi); b),
then the log-likelihood contribution from presence site i
for b is
 ψ( y  1z ( x ); β) 
log  i i
  Ci
 ∑ x ∈χ ψ ( y  1z ( x ); β) 
(15)
where Ci can be discarded, since it does not depend on b.
867